Almost the same exept 2 million polys. Render time roughly the same as before. Still don’t understand what does memory bandwith have to do with how fast the path tracing is.
If you were right then the render times using Redshift with the same scene without the statues would be similar. But here is my times with Redshift 2.6.32 demo:
I was about to buy an ngreedia rtx 2070 but fortunately i change my mind since amd hire a programmer to get opencl correctly implemented. Now i am just waiting for a good amd card that will probably be only available in fall.
I’m just not seeing it.
You rendered an image with a peak memory of 38.03mb in 7 minutes. You then rendered a scene with a peak memory of 461.69mb in the same amount of time using the same card. That shows the original demo scene was not enough of a stress test to get a proper read, and that the 2060 can handle a much larger bandwidth.
I just tried to prove you were wrong. By the way the peak memory usage of the BMW scene is 144MB.
If you were right, I would get similar results with the same scene in Redshift. But that was not the case, see my previous reply above with the render times.
Then why is that if you were right? Does Redshift handle something that wrong that the render times will be better?
I’m not going to talk about Redshift because I’ve never used it and am unfamiliar with it. However, none of that changes the original point I tried to make at the beginning of this thread; That your original demo scene is NOT a good test subject.
It’s the only scene in which you are getting inconsistent results. All of the (proper) benchmark scenes are consistently showing faster times with the 2060 vs the 1060. I even said previously that you could be right about a potential issue with the card, so it’s not like I’m just trying to prove you wrong. I’m just saying that you need to base your finding on a proper benchmark and ditch your demo scene.
I agree that memory bandwidth impact should be accounted for in the expected speedup. Looking at FLOPS (pure floating-point computation), the 2060 is only about 35% faster than the 1060.
Also, Cycles often initially performs worse with new major CUDA SDK releases (required for new GPUs) and then gets better over time.
No.
Path tracing uses memory to store geometry, obviously. If your scene has very little geometry, a lot of that will end up in fast on-chip caches, so memory bandwidth becomes less relevant.
You’re comparing apples and oranges. Unlike Cycles, Redshift uses light caches which also occupy memory. Memory bandwidth becomes more important.
Also unlike Cycles, most renderers these days don’t use unikernels anymore, which also leads to higher bandwidth usage.
I made a new scene from scratch based on the previous one. I’m still confused what might be the problem with a such simple scene. I didn’t use anything from earlier. All properties are the same: samples, max bounces…etc.
And look at those crazy render time: 16:08. And maybe it’s not just RTX related.